Files
paddlepaddle--paddlenlp/paddlenlp/transformers/tensor_parallel_utils.py
T
wehub-resource-sync 2aaeece67c
Codestyle Check / Lint (push) Has been cancelled
Codestyle Check / Check bypass (push) Has been cancelled
Pipelines-Test / Pipelines-Test (push) Has been cancelled
chore: import upstream snapshot with attribution
2026-07-13 13:37:14 +08:00

584 lines
23 KiB
Python

# Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import paddle
import paddle.distributed.fleet as fleet
try:
from paddle.nn.layer.layers import in_declarative_mode
except:
from paddle.fluid.dygraph.base import in_declarative_mode
import paddle.distributed as dist
from paddle.autograd import PyLayer
from paddlenlp.utils.tools import get_env_device
def parallel_matmul(lm_output, logit_weights, tensor_parallel_output=True, training=True):
"""
Parallel matmul
Args:
lm_output: x for matmul
logit_weights: y for matmul
tensor_parallel_output: the output is paralleled or not
training: args for xpu
Returns:
rst for matmul
"""
if get_env_device() == "xpu":
try:
from paddle_xpu.layers.nn import parallel_matmul as xpu_parallel_matmul
xpu_parallel_matmul = xpu_parallel_matmul()
logits = xpu_parallel_matmul(
lm_output,
logit_weights,
tensor_parallel_output=tensor_parallel_output,
training=training,
)
return logits
except ImportError:
pass
is_fleet_init = True
tensor_parallel_degree = 1
try:
hcg = fleet.get_hybrid_communicate_group()
model_parallel_group = hcg.get_model_parallel_group()
tensor_parallel_degree = hcg.get_model_parallel_world_size()
except:
is_fleet_init = False
is_logit_weight_distributed = logit_weights.is_distributed
# `is_distributed` in static mode is always False
if in_declarative_mode() and tensor_parallel_degree > 1:
is_logit_weight_distributed = True
if is_fleet_init and tensor_parallel_degree > 1 and is_logit_weight_distributed:
input_parallel = paddle.distributed.collective._c_identity(lm_output, group=model_parallel_group)
logits = paddle.matmul(input_parallel, logit_weights, transpose_y=True)
if tensor_parallel_output:
return logits
return paddle.distributed.collective._c_concat(logits, group=model_parallel_group)
else:
logits = paddle.matmul(lm_output, logit_weights, transpose_y=True)
return logits
def parallel_linear(lm_output, logit_weights, bias, tensor_parallel_output=True):
is_fleet_init = True
tensor_parallel_degree = 1
try:
hcg = fleet.get_hybrid_communicate_group()
model_parallel_group = hcg.get_model_parallel_group()
tensor_parallel_degree = hcg.get_model_parallel_world_size()
except:
is_fleet_init = False
is_logit_weight_distributed = logit_weights.is_distributed
# `is_distributed` in static mode is always False
if in_declarative_mode() and tensor_parallel_degree > 1:
is_logit_weight_distributed = True
if is_fleet_init and tensor_parallel_degree > 1 and is_logit_weight_distributed:
input_parallel = paddle.distributed.collective._c_identity(lm_output, group=model_parallel_group)
bias_parallel = paddle.distributed.collective._c_identity(bias, group=model_parallel_group)
logits = paddle.matmul(input_parallel, logit_weights)
logits += bias_parallel
if tensor_parallel_output:
return logits
return paddle.distributed.collective._c_concat(logits, group=model_parallel_group)
else:
logits = paddle.matmul(lm_output, logit_weights)
logits += bias
return logits
def fused_head_and_loss_fn(
hidden_states,
lm_head_weight,
lm_head_bias,
labels,
loss_mask,
transpose_y,
num_embeddings,
tensor_parallel_degree,
tensor_parallel_output,
fused_linear,
loop_chunk_size,
return_token_loss,
ignore_index,
):
"""Run FusedHeadAndCrossEntropy."""
return FusedHeadAndCrossEntropy.apply(
hidden_states,
lm_head_weight,
lm_head_bias,
labels,
loss_mask,
transpose_y,
num_embeddings,
tensor_parallel_degree,
tensor_parallel_output,
fused_linear,
loop_chunk_size,
return_token_loss,
ignore_index,
)
class FusedHeadAndCrossEntropy(PyLayer):
"""Fuse LM Head and CrossEntropyLoss into one module."""
@staticmethod
def forward(
ctx,
hidden_states: paddle.Tensor,
lm_head_weight: paddle.Tensor,
lm_head_bias: paddle.Tensor,
labels: paddle.Tensor,
loss_mask: paddle.Tensor,
transpose_y: bool,
num_embeddings: int,
tensor_parallel_degree: int,
tensor_parallel_output: bool,
fused_linear: bool,
loop_chunk_size: int,
return_token_loss: bool,
ignore_index: int,
):
"""Run blockwise parallel cross entropy calculation.
Args:
ctx: PyLayerContext
hidden_states (`paddle.Tensor` of shape `(batch_size, max_seq_len, hidden_size)`): the input features.
lm_head_weight (`paddle.Tensor` of shape `(hidden_size, vocab_size)`)
lm_head_bias (`paddle.Tensor` of shape `(vocab_size)`)
labels (`paddle.Tensor` of shape `(batch_size, max_seq_len)`)
loss_mask (`paddle.Tensor` of shape `(batch_size, max_seq_len)`)
transpose_y: bool
num_embeddings: int
tensor_parallel_degree: int
tensor_parallel_output: bool
fused_linear: bool
loop_chunk_size: int, default is LOOP_CHUNK_SIZE
return_token_loss: bool
ignore_index: int
Returns:
loss (`paddle.Tensor` of shape `()`: the output loss.
"""
if fused_linear:
# print("Cannot support fused_linear while using use_fused_head_and_loss_fn now!")
fused_linear = False # NOTE(hehuang): Cannot support fused_linear now
# initialize distributed settings
dtype = hidden_states.dtype
if tensor_parallel_degree > 1:
hcg = fleet.get_hybrid_communicate_group()
model_parallel_group = hcg.get_model_parallel_group()
tensor_parallel_degree = hcg.get_model_parallel_world_size()
original_shape = hidden_states.shape
hidden_states = hidden_states.reshape([-1, original_shape[-1]])
labels = labels.reshape([-1])
if loss_mask is None:
ctx.aux_num = 1
loss_mask = (labels != ignore_index).astype("float32")
else:
ctx.aux_num = 2
loss_mask = loss_mask.reshape([-1]).astype("float32")
ctx.return_token_loss = return_token_loss
if return_token_loss:
divisor = 1
else:
divisor = loss_mask.sum()
n_tokens = hidden_states.shape[0]
n_classes = lm_head_weight.shape[0] if transpose_y else lm_head_weight.shape[1]
# cast lm_head weight & bias
lm_head_weight_cast = lm_head_weight.astype(dtype)
if lm_head_bias is not None:
lm_head_bias_cast = lm_head_bias.astype(dtype)
# initialize indices for labels_one_hot
if tensor_parallel_degree > 1 and tensor_parallel_output:
rank = hcg.get_model_parallel_rank()
per_part_size = num_embeddings // tensor_parallel_degree
indices = paddle.arange(
rank * per_part_size,
rank * per_part_size + n_classes,
dtype=labels.dtype,
).unsqueeze(0)
else:
indices = paddle.arange(num_embeddings, dtype=labels.dtype).unsqueeze(0)
# initialize gradients
if not return_token_loss:
if not lm_head_weight.stop_gradient:
grad_lm_head_weight = paddle.zeros_like(lm_head_weight)
else:
grad_lm_head_weight = None
if lm_head_weight is not None and not lm_head_weight.stop_gradient:
grad_lm_head_bias = paddle.zeros_like(lm_head_bias)
else:
grad_lm_head_bias = None
if hidden_states.stop_gradient:
grad_hidden_states = paddle.zeros_like(hidden_states)
else:
grad_hidden_states = None
# initialize outputs
token_loss = paddle.empty((n_tokens,), dtype=paddle.float32)
# blockwise calculations
for i in range(0, n_tokens, loop_chunk_size):
token_start_idx = i
token_end_idx = min(i + loop_chunk_size, n_tokens)
cur_chunk_range = paddle.arange(token_start_idx, token_end_idx)
hidden_states_chunk = paddle.gather(hidden_states, cur_chunk_range, axis=0)
labels_chunk = paddle.gather(labels, cur_chunk_range, axis=0)
loss_mask_chunk = paddle.gather(loss_mask, cur_chunk_range, axis=0)
# logits calculations
logits_chunk_cast = paddle.matmul(
hidden_states_chunk,
lm_head_weight_cast,
transpose_y=transpose_y,
)
if lm_head_bias is not None:
logits_chunk_cast += lm_head_bias_cast
if tensor_parallel_degree > 1 and not tensor_parallel_output:
logits_chunk_cast_lst = []
dist.all_gather(
logits_chunk_cast_lst,
logits_chunk_cast,
group=model_parallel_group,
)
logits_chunk_cast = paddle.concat(logits_chunk_cast_lst, axis=-1)
logits_chunk = logits_chunk_cast.astype("float32")
# log softmax
max_logits = paddle.max(logits_chunk, axis=-1, keepdim=True)
if tensor_parallel_degree > 1 and tensor_parallel_output:
dist.all_reduce(max_logits, op=dist.ReduceOp.MAX, group=model_parallel_group)
normalized_logits = logits_chunk - max_logits
exp_logits = paddle.exp(normalized_logits)
sum_exp_logits = paddle.sum(exp_logits, axis=-1, keepdim=True)
if tensor_parallel_degree > 1 and tensor_parallel_output:
dist.all_reduce(
sum_exp_logits,
op=dist.ReduceOp.SUM,
group=model_parallel_group,
)
log_sum_exp_logits = paddle.log(sum_exp_logits)
# cross entropy
labels_one_hot = labels_chunk.unsqueeze(1) == indices
label_logits = paddle.sum(
paddle.where(
labels_one_hot,
normalized_logits,
paddle.zeros_like(normalized_logits),
),
axis=-1,
keepdim=True,
)
if tensor_parallel_degree > 1 and tensor_parallel_output:
dist.all_reduce(
label_logits,
op=dist.ReduceOp.SUM,
group=model_parallel_group,
)
token_loss_chunk = (log_sum_exp_logits - label_logits).squeeze(1) / divisor
cond = loss_mask_chunk.astype("bool")
token_loss_chunk = paddle.where(cond, token_loss_chunk, paddle.zeros_like(token_loss_chunk))
paddle.scatter_(token_loss, cur_chunk_range, token_loss_chunk, overwrite=True)
# gradients calculations
if not return_token_loss:
if tensor_parallel_degree > 1 and not tensor_parallel_output:
exp_logits = exp_logits.split(model_parallel_group.nranks, axis=-1)[model_parallel_group.rank]
labels_one_hot = labels_one_hot.split(model_parallel_group.nranks, axis=-1)[
model_parallel_group.rank
]
grad_logits_chunk = (exp_logits / sum_exp_logits - labels_one_hot.astype("float32")) / divisor
grad_logits_chunk = grad_logits_chunk.astype(dtype)
grad_logits_chunk = paddle.where(
cond.unsqueeze(1),
grad_logits_chunk,
paddle.zeros_like(grad_logits_chunk),
)
if grad_hidden_states is not None:
paddle.scatter_(
grad_hidden_states,
cur_chunk_range,
paddle.matmul(grad_logits_chunk, lm_head_weight_cast, transpose_y=not transpose_y),
overwrite=True,
)
if grad_lm_head_weight is not None:
if transpose_y:
grad_lm_head_weight += paddle.matmul(
grad_logits_chunk,
hidden_states_chunk,
transpose_x=True,
)
else:
grad_lm_head_weight += paddle.matmul(
hidden_states_chunk,
grad_logits_chunk,
transpose_x=True,
)
if grad_lm_head_bias is not None:
grad_lm_head_bias += grad_logits_chunk.astype("float32").sum(axis=0).astype(dtype)
if return_token_loss:
loss = token_loss.reshape(original_shape[:-1])
ctx.save_for_backward(
hidden_states,
lm_head_weight,
lm_head_bias,
labels,
loss_mask,
)
ctx.transpose_y = transpose_y
ctx.num_embeddings = num_embeddings
ctx.loop_chunk_size = loop_chunk_size
ctx.tensor_parallel_degree = tensor_parallel_degree
ctx.tensor_parallel_output = tensor_parallel_output
ctx.original_shape = original_shape
else:
loss = token_loss.sum()
ctx.hidden_states_has_grad = grad_hidden_states is not None
ctx.lm_head_weight_has_grad = grad_lm_head_weight is not None
ctx.lm_head_bias_has_grad = grad_lm_head_bias is not None
grad_args = []
if ctx.hidden_states_has_grad:
if tensor_parallel_degree > 1:
dist.all_reduce(
grad_hidden_states,
op=dist.ReduceOp.SUM,
group=model_parallel_group,
)
grad_args.append(grad_hidden_states.reshape(original_shape))
if ctx.lm_head_weight_has_grad:
grad_args.append(grad_lm_head_weight)
if ctx.lm_head_bias_has_grad:
grad_args.append(grad_lm_head_bias)
ctx.save_for_backward(*grad_args)
return loss
@staticmethod
def backward(ctx, grad_output):
"""Run the backward of blockwise parallel cross entropy calculation."""
if not ctx.return_token_loss:
grad_args = ctx.saved_tensor()
idx = 0
if ctx.hidden_states_has_grad:
grad_hidden_states = grad_args[idx] * grad_output.astype(grad_args[idx].dtype)
idx += 1
else:
grad_hidden_states = None
if ctx.lm_head_weight_has_grad:
grad_lm_head_weight = grad_args[idx] * grad_output.astype(grad_args[idx].dtype)
idx += 1
else:
grad_lm_head_weight = None
if ctx.lm_head_bias_has_grad:
grad_lm_head_bias = grad_args[idx] * grad_output.astype(grad_args[idx].dtype)
idx += 1
else:
grad_lm_head_bias = None
if ctx.aux_num == 1:
return (
grad_hidden_states,
grad_lm_head_weight,
grad_lm_head_bias,
None,
)
else:
return (
grad_hidden_states,
grad_lm_head_weight,
grad_lm_head_bias,
None,
None,
)
# return_token_loss = True
grad_token_loss = grad_output.reshape([-1])
(
hidden_states,
lm_head_weight,
lm_head_bias,
labels,
loss_mask,
) = ctx.saved_tensor()
transpose_y = ctx.transpose_y
num_embeddings = ctx.num_embeddings
loop_chunk_size = ctx.loop_chunk_size
tensor_parallel_degree = ctx.tensor_parallel_degree
tensor_parallel_output = ctx.tensor_parallel_output
# initialize distributed settings
dtype = hidden_states.dtype
if tensor_parallel_degree > 1:
hcg = fleet.get_hybrid_communicate_group()
model_parallel_group = hcg.get_model_parallel_group()
tensor_parallel_degree = hcg.get_model_parallel_world_size()
n_tokens = hidden_states.shape[0]
n_classes = lm_head_weight.shape[0] if transpose_y else lm_head_weight.shape[1]
# cast lm_head weight & bias
lm_head_weight_cast = lm_head_weight.astype(dtype)
if lm_head_bias is not None:
lm_head_bias_cast = lm_head_bias.astype(dtype)
# initialize indices for labels_one_hot
if tensor_parallel_degree > 1 and tensor_parallel_output:
rank = hcg.get_model_parallel_rank()
per_part_size = num_embeddings // tensor_parallel_degree
indices = paddle.arange(
rank * per_part_size,
rank * per_part_size + n_classes,
dtype=labels.dtype,
).unsqueeze(0)
else:
indices = paddle.arange(num_embeddings, dtype=labels.dtype).unsqueeze(0)
# initialize gradients
if not lm_head_weight.stop_gradient:
grad_lm_head_weight = paddle.zeros_like(lm_head_weight)
else:
grad_lm_head_weight = None
if lm_head_weight is not None and not lm_head_weight.stop_gradient:
grad_lm_head_bias = paddle.zeros_like(lm_head_bias)
else:
grad_lm_head_bias = None
if hidden_states.stop_gradient:
grad_hidden_states = paddle.zeros_like(hidden_states)
else:
grad_hidden_states = None
# blockwise calculations
for i in range(0, n_tokens, loop_chunk_size):
token_start_idx = i
token_end_idx = min(i + loop_chunk_size, n_tokens)
cur_chunk_range = paddle.arange(token_start_idx, token_end_idx)
hidden_states_chunk = paddle.gather(hidden_states, cur_chunk_range, axis=0)
labels_chunk = paddle.gather(labels, cur_chunk_range, axis=0)
loss_mask_chunk = paddle.gather(loss_mask, cur_chunk_range, axis=0)
# logits calculations
logits_chunk_cast = paddle.matmul(
hidden_states_chunk,
lm_head_weight_cast,
transpose_y=transpose_y,
)
if lm_head_bias is not None:
logits_chunk_cast += lm_head_bias_cast
if tensor_parallel_degree > 1 and not tensor_parallel_output:
logits_chunk_cast_lst = []
dist.all_gather(
logits_chunk_cast_lst,
logits_chunk_cast,
group=model_parallel_group,
)
logits_chunk_cast = paddle.concat(logits_chunk_cast_lst, axis=-1)
logits_chunk = logits_chunk_cast.astype("float32")
# log softmax
max_logits = paddle.max(logits_chunk, axis=-1, keepdim=True)
if tensor_parallel_degree > 1 and tensor_parallel_output:
dist.all_reduce(max_logits, op=dist.ReduceOp.MAX, group=model_parallel_group)
normalized_logits = logits_chunk - max_logits
exp_logits = paddle.exp(normalized_logits)
sum_exp_logits = paddle.sum(exp_logits, axis=-1, keepdim=True)
if tensor_parallel_degree > 1 and tensor_parallel_output:
dist.all_reduce(
sum_exp_logits,
op=dist.ReduceOp.SUM,
group=model_parallel_group,
)
labels_one_hot = labels_chunk.unsqueeze(1) == indices
if tensor_parallel_degree > 1 and not tensor_parallel_output:
exp_logits = exp_logits.split(model_parallel_group.nranks, axis=-1)[model_parallel_group.rank]
labels_one_hot = labels_one_hot.split(model_parallel_group.nranks, axis=-1)[model_parallel_group.rank]
grad_logits_chunk = exp_logits / sum_exp_logits - labels_one_hot.astype("float32")
# NOTE(hehuang): scaling grad_logits_chunk by grad_token_loss
grad_logits_chunk *= paddle.gather(grad_token_loss, cur_chunk_range, axis=0).unsqueeze(1)
grad_logits_chunk = grad_logits_chunk.astype(dtype)
cond = loss_mask_chunk.astype("bool")
grad_logits_chunk = paddle.where(
cond.unsqueeze(1),
grad_logits_chunk,
paddle.zeros_like(grad_logits_chunk),
)
if grad_hidden_states is not None:
paddle.scatter_(
grad_hidden_states,
cur_chunk_range,
paddle.matmul(grad_logits_chunk, lm_head_weight_cast, transpose_y=not transpose_y),
overwrite=True,
)
if grad_lm_head_weight is not None:
if transpose_y:
grad_lm_head_weight += paddle.matmul(grad_logits_chunk, hidden_states_chunk, transpose_x=True)
else:
grad_lm_head_weight += paddle.matmul(hidden_states_chunk, grad_logits_chunk, transpose_x=True)
if grad_lm_head_bias is not None:
grad_lm_head_bias += grad_logits_chunk.astype("float32").sum(axis=0).astype(dtype)
if grad_hidden_states is not None:
if tensor_parallel_degree > 1:
dist.all_reduce(
grad_hidden_states,
op=dist.ReduceOp.SUM,
group=model_parallel_group,
)
grad_hidden_states = grad_hidden_states.reshape(ctx.original_shape)
if ctx.aux_num == 1:
return (
grad_hidden_states,
grad_lm_head_weight,
grad_lm_head_bias,
None,
)
else:
return (
grad_hidden_states,
grad_lm_head_weight,
grad_lm_head_bias,
None,
None,
)